IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 60, NO.
11, NOVEMBER 2015
Optimal Denial-of-Service Attack Scheduling
With Energy Constraint
Heng Zhang, Peng Cheng, Ling Shi, and Jiming Chen
Abstract—Security of Cyber-Physical Systems (CPS) has gained
increasing attention in recent years. Most existing works mainly
investigate the system performance given some attacking patterns.
In this technical note, we investigate how an attacker should
schedule its Denial-of-Service (DoS) attacks to degrade the system
performance. Specifically, we consider the scenario where a sensor
sends its data to a remote estimator through a wireless channel,
while an energy-constrained attacker decides whether to jam
the channel at each sampling time. We construct optimal attack
schedules to maximize the expected average estimation error at
the remote estimator. We also provide the optimal attack schedules
when a special intrusion detection system (IDS) at the estimator is
given. We further discuss the optimal attack schedules when the
sensor has energy constraint. Numerical examples are presented
to demonstrate the effectiveness of the proposed optimal attack
schedules.
Index Terms—DoS attack, energy constraint, state estimation.
I. I NTRODUCTION
Recently, Cyber-Physical Systems (CPS), in which information and
physical elements are closely integrated, have gained much research
interest from various aspects [3]–[5]. Applications of CPS have a wide
spectrum, such as battle field, smart grid, smart building, intelligent
transportation, etc. CPS are vulnerable to different types of malicious
attacks which make the security a fundamental issue for reliable
services [3]. A famous example is that in June 2010 a cyber worm
named “stuxnet” attacked an Iranian nuclear facility at Natanz, which
resulted in 60% hosts damaged [6].
Researchers have investigated security issues in CPS from different
perspectives recently [7], [8]. Teixeira et al. [8] summarized the
characteristics of network attacks and defined an attack space with the
adversary’s priori knowledge of the system model, its disclosure, and
disruption resources. Typical attacks can be categorized into Denial-
of-Service (DoS) attack, replay data attack, and false data injection
attack [8]. DoS attack that blocks the communication between the sys-
tem components has been well studied, since it is the most reachable
Manuscript received June 9, 2014; revised October 19, 2014, February 1,
2015, and February 11, 2015; accepted February 14, 2015. Date of publication
March 4, 2015; date of current version October 26, 2015. This work was
supported in part by the NSFC under Grant U1401253, the National Program
for Special Support of Top-Notch Young Professionals, Fundamental Research
Funds for the Central Universities 2014XZZX003-25, NCET-11-0445, and by
an HKUST Caltech Partnership FP004. This work was presented in part at the
IEEE CDC 2013, Florence, Italy, December 2013. Recommended by Associate
Editor W. X. Zheng. (Corresponding author: Peng Cheng.)
H. Zhang, P. Cheng, and J. Chen are with State Key Laboratory of In-
dustrial Control Technology, Zhejiang University, Hangzhou, China (e-mail:
ezhangheng@gmail.com; pcheng@iipc.zju.edu.cn; jmchen@iipc.zju.edu.cn).
L. Shi is with the Department of Electronic and Computer Engineering,
Hong Kong University of Science and Technology, Hong Kong, China (e-mail:
eesling@ust.hk).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TAC.2015.2409905
0018-9286 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Thapar Institute of Engineering & Technology. Downloaded on March 15,2023 at 18:23:08 UTC from IEEE Xplore. Restrictions apply.
3023
attack pattern in the attack space [9]. Besides the theoretical research
on DoS attack, some experiments were also conducted to demonstrate
the effect of DoS attacks [10].
In practice, energy constraint is a natural concern for various types
of attackers, which will affect their attack policies [11]–[13]. Law et al.
[11] pointed out that jammers may run out of energy very fast when
their energy budget is limited. Kavitha et al. [12] introduced a random
DoS attacker who randomly decides to jam the channel or sleep in
order to save the energy.
Most existing works mainly investigate the system performance
for given attack patterns. The study of the adversary’s optimal attack
schedules and the corresponding consequences is an important aspect
in CPS security as pointed out in [14]. Motivated by this, we aim
to answer how to optimally schedule the attack so as to maximize
the attacking effect on a CPS. Specifically, we consider a system
where one sensor processes the measurements and sends the data
to a remote estimator through a wireless channel. An attacker has a
limited attacking energy budget, and decides at each sampling time
whether or not to jam the channel in order to degrade the remote
estimation quality. To the best of our knowledge, this is the first work
on the DoS attack schedules against remote state estimation. The main
contributions of this technical note are summarized as follows:
1) We construct the optimal DoS attack schedules, which maximize
the expected average estimation error.
2) We present the optimal attack schedules when a special IDS in
the estimator is given.
3) We further present the optimal attack schedules when both the
sensor and the attacker have energy constraints.
The remainder of the technical note is organized as follows. In
Section II, we formulate the attack schedule problems. In Section III,
we construct the optimal DoS attack schedules for maximizing the ex-
pected average estimation error. In Section IV, we present the optimal
attack schedules for avoiding a given intrusion detection system. In
Section V, we study the optimal attack schedules when the sensor also
has energy constraint. In Section VI, numerical examples are shown to
illustrate the results. Finally, Section VII concludes the technical note.
Notations: Sn + stands for the set of n × n positive semi-definite
matrices. Z+ stands for the set of positive integers. E[X] is the mean
of random variable X, and E[X|Y ] is the mean of random variable X
conditioned on Y , respectively. γ0 is used to denote the zero norm
of γ, which is the number of nonzero entries of the vector γ. T r(·)
represents the trace of matrix.
II. P ROBLEM F ORMULATION AND P RELIMINARIES
A. Problem Formulation
Consider the following system (Fig. 1):
xk+1 = Axk + wk
yk = Cxk + vk (1)
where xk ∈ Rnx is the system state with nx ∈ Z+ , yk ∈ Rny is the
sensor measurement with ny ∈ Z+ , wk ∈ Rnx is the process noise,
vk ∈ Rny is the measurement noise, and wk and vk are uncorrelated
 3024 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 60, NO. 11, NOVEMBER 2015

P , then it can be easily obtained that Pks = P for all k ∈ [1, T ].

 Δ Δ
Define h, hk : Sn + → S+ as h(X) = AXA + Σw , and h (X) =
x nx k

h
 ◦ h ◦
· · · ◦ h
 (X).
k times
From [17], the following result holds.
Lemma 3.1: The function h has following property:
Fig. 1. Schematic of attack on the network. P ≤ h(P ) ≤ h2 (P ) ≤ · · · ≤ hk (P ) ≤ · · · , ∀k ∈ Z+ .
zero mean Gaussian noises with covariance Σw and Σv , respec-
1/2 It is straightforward to show that at the remote estimator, x̂k and Pk
tively. The pair (A, C) is assumed to be observable and (A, Σw ) is
are obtained as follows [18]:
controllable. 
After yk is obtained, the sensor pre-estimates the state xk , and (Ax̂ k−1 , h(Pk−1 )) , if γk = 1 and θk = 1;
then sends its minimum mean squared error (MMSE) estimate x̂sk = (x̂k , Pk ) = s
x̂k , P , otherwise.
E[xk |y1 , . . . , yk ] to a remote estimator through a wireless channel.
We consider the scenario that there is an attacker which degrades Assume that the data packet is delivered to the remote estimator
the remote estimation quality by jamming the wireless channel. The successfully at time s, and the attacker launches m consecutive attacks
attacker has a limited energy budget and has to determine whether to from time s + 1 to s + m. The state space of the error covariance
jam the channel or not at each sampling time. It is assumed that the matrix Pk in this period is {P , h(P ), . . . , hm (P )}.
packet will arrive at the remote estimator if there is no DoS attack We can see that the error covariance matrix Pk during the consecu-
during the transmission. tive attack period [s + 1, s + m] is a stationary Markov chain, and the
Denote γk = 1 or 0 as the attacker’s decision variable at time k, transition probability matrix (pij ) is given by
and γ as a schedule that defines γk at each k. When the attacker α, j = i+1;
launches an attack and jams the channel, the sensor data packet will pij = P r Pk = hj (P )|Pk−1 = hi (P ) = 1−α, j = 0; (2)
be dropped with probability α. Denote θk = 1 or 0 as the indicator 0, others
function whether the data packet drops or not. We assume that θk ’s are
i.i.d. Bernoulli random variables with E(θk ) = α.1 i, j = 0, 1, 2, . . . , m. Moreover, we have the following properties for
Denote all the data packets received at the estimator until time Pk , k ∈ [s + 1, s + m].
k as D(γ1:k ). The remote estimator then computes its own MMSE Property 3.1: During the consecutive attack period [s + 1, s + m]:
estimate x̂k and its corresponding estimation error covariance Pk 1) the distribution of the error covariance matrix Ps+k can be
based on D(γ1:k ), i.e., x̂k (γ1:k ) = E[xk |D(γ1:k )] and Pk (γ1:k ) = computed as
E[(xk − x̂k )(xk − x̂k ) |D(γ1:k )]. For simplicity, we write x̂k (γ1:k ) 
αi − αi+1 , i = 0, 1, . . . , k − 1;
as x̂k , etc., when the schedule γ1:k is given. P r Ps+k = hi (P ) =
αk , i = k.
Consider a finite horizon T . Due to the finite energy constraint, we
assume that the attacker can only launch n attacks from k = 1 to k = 2) the conditional probability can be computed as follows:
T , i.e., γ0 = n. 
Average Error: For a given attack schedule γ, define Ja (γ) as the αj − αj+1 , j = 0, . . . , k−1;
P r Ps+k = hj (P )|Ps = hi (P ) =
average expected estimation error covariance matrix, i.e., αk , j = i + k.

1 
T
This property can be easily obtained from (2) by mathematical induc-
Ja (γ) = E [Pk (γk )] . tion method. Due to the space limitation, the proof is omitted.
T
k=1 Property 3.2: 1) If the packet is not received at the remote estimator
In this technical note, we aim to solve the following problem in the at time s, then E[Ps+m |Ps ≥ P ] ≥ E[Ps+m |Ps = P ].
viewpoint of the attacker: 2) Consider two different consecutive attacking periods, n1 and n2 ,
Problem 2.1: where n1 ≥ n2 . Then E[Ps+n1 ] ≥ E[Ps+n2 ].
An attack scheme with given attacking number n can be divided into
max T r [Ja (γ)] the following consecutive attacking sequences, k1 , k2 , . . . , ks , i.e.,
γ∈Γ
s.t. γ0 = n, (0, · · ·, 0, 1, · · ·, 1, 0, · · ·, 0, 1, · · ·, 1, 0, · · ·, 0, 1, · · ·, 1, 0, · · ·, 0),
        
where Γ = {0, 1}T is the set of all possible attack schedules. k1 times k2 times ks times

s
where ki = n. Note that each neighboring sequences are di-
i=1
III. O PTIMAL ATTACK S CHEDULE A NALYSIS vided by at least one zero. Using Property 3.1, we have
A. State Estimation Under DoS Attack
1 1 
T s ki
T −n
In this section, we present some properties of the estimation error at Ja (γ) = E [Pk (γk )] = E [Psi +ni ]+ P
T T T
the remote estimator under a DoS attack. k=1 i=1 ni =1

The MMSE estimate x̂sk is obtained from a standard Kalman n −1 

1 
s ki i
filter. Let Pks be the estimation error covariance matrix of x̂sk . Stan- T −n
= h (P )(α −α
m m m+1 ni
)+α h (P ) + ni
P.
dard Kalman filtering analysis shows that Pks converges exponen- T T
i=1 ni =1 m=0
tially to its steady-state value P [16]. For brevity, we assume Π0 =
(3)
1 Since the bit error rate (BER) is fixed in every attack time when the attacker
exploits the same energy to jam the channel [15], we assume that the rate of Note that (3) is independent of the attack start sequence si , i =
successful attacks follows a Bernoulli probability distribution. 1, 2, . . . , s, which means that any permutation of the ki ’s will lead to

Authorized licensed use limited to: Thapar Institute of Engineering & Technology. Downloaded on March 15,2023 at 18:23:08 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 60, NO. 11, NOVEMBER 2015 3025

the same average expected estimation error. Thus, for the consecutive
k1 ⊕k2 ⊕···⊕ks
attack sequences k1 , k2 , . . . , ks , we define γ1:T as the set of
(k1 ⊕k2 ⊕···⊕ks )
attack schedules which have the same performance Ja,1:T .
k1 ⊕k2 ⊕···⊕ks k1 ⊕k2 ⊕···⊕ks
For simplicity, we write γ1:T as γ , and write
(k1 ⊕k2 ⊕···⊕ks ) (k ⊕k ⊕···⊕ks )
Ja,1:T as Ja 1 2 , when the time horizon [1, T ] is
given. In particular, let γ n be the set of attack schedules with a con-
(n)
secutive attack sequence n, which have the same performance Ja .
From Property 3.2, we have following property. Fig. 2. Schematic of attack on the network against state estimation with an
Property 3.3: The following statements are true. intrusion detector.
(n ) (n ) 1, m2 = m − m1 if m is odd, with m = z + [n mod (T − n)].3 The
1) Ja 1 ≤ Ja 2 , where n1 ≤ n2 .
(n ⊕n ) (n) resulting cost is
2) Ja 1 2 ≤ Ja , where n = n1 + n2 .
(n1 ⊕n2 ⊕···⊕ns ) (n)
3) Ja ≤ Ja , where n = n1 + n2 + · · · + ns . 
z

m1
(m1 ⊕m2 ) (n1 ⊕n2 ) 1
4) Ja ≤ Ja , where m1 + m2 = n1 + n2 and T r(Ja )min = Tr (T −n− 1) hi (α, P ) + hi (α, P )
T
max{m1 , m2 , n1 , n2 } is n1 or n2 . i=1 i=1

m2
From statements 1) and 4) in Property 3.3, we can see that the + hi (α, P ) + [(T − n) + m(1 − α)
more the attack times are grouped together, the lager the system cost i=1
becomes. Statements 2) and 3) demonstrate that consecutive attack is + z(1 − α)(T − n − 1)] P , (6)
much better than scattered attack from the viewpoint of attacker.
where hi (α, P ) = [(n − i)(αi − αi+1 ) + αi ]hi (P ).
Proof: A direct result from Property 3.3 3) and 4). 
B. Optimal Attack Schedule Design From the above two theorems, one can see that grouping the attacks
Now we are ready to present the optimal attack schedule for together leads to maximal attacking effect, while separating the attacks
Problem 2.1. as uniformly as possible leads to the minimal degradation. Thus the
Theorem 3.1: The optimal attack schedule that solves Problem 2.1 results can be viewed as the upper and lower bound of the damage an
is anyone that belongs to γ n , and the corresponding cost is given by arbitrary attack schedule can make.

1    T − nα
n

T r(Ja )max = T r hi (α, P ) + T r(P ), (4) IV. O PTIMAL ATTACK S CHEDULES AGAINST A G IVEN IDS
T T
i=1 The aforementioned problems are discussed under the assumption
that the channel is perfect and there is no defensive measure. In
where hi (α, P ) = [(n − i)(αi − αi+1 ) + αi ]hi (P ). practical applications, when the estimator considers unknown factors
Proof: A direct result from Property 3.3 2) and 3).  from environment changes or adversary’s intrusion, which may lead to
As a byproduct, we can obtain the worst attack strategy which data drops, a detector is often designed at the receiver side as a first
minimizes the expected average error covariance. step to reduce the effectiveness of these factors (Fig. 2). For example,
Problem 3.1: packet reception rate (P RR) at the receiver side is proposed as the
criteria for intrusion detection in [19]. P RR is the rate of packets
min T r [Ja (γ)] reception that are successfully delivered to the estimator compared to
γ∈Γ
the number of packets that have been sent out by the sensor.
s.t. γ0 = n. In practice, the P RR can be computed for a given length of a
time window τ , i.e., P RR = σ/τ , where σ is the packet reception
The solution of Problem 3.1 is given by the following theorem. numbers in the time window. If the P RR is too small, the detector
Theorem 3.2: 1) If n ≤ (1/2)T , an optimal solution for will deduce that there may exist an attacker, which triggers an alarm to
Problem 3.1 is anyone in γ 1⊕1⊕···⊕1 , and the corresponding cost is alert the sensor. Then the sensor can use channel hopping technology
given by to guarantee the packet can be received by the remote estimator [20].
In this technical note, we assume that the alarm is triggered if P RR ≤
T − nα nα   P RR0 , where P RR0 is a given alarm bound.4
T r(Ja )min = T r(P ) + T r h(P ) . (5)
T T Before introducing the attack schedule against the IDS mecha-
nism, we present a property to help formulate the attack optimization
2) If n > (1/2)T , an optimal solution for Problem 3.1 is problem.
Property 4.1: The alarm triggering condition P RR ≤ P RR0 is
(1 · · 1 0) · · · (1
 ·  ·
··1
 0)( 1 ·
· · 1 0 1
 ·
· · 1 ) (0 1 ·
· · 1) · · · (0 1 ·
· · 1) equivalent to Pk  hd (P ), k = 1, 2, . . . , T , where d = τ − σ0 , σ0 =
max{σ|(σ/τ ) ≤ P RR0 }.
     
z times z times m1 times m2 times z times z times
Proof: See the Appendix. 
r1 times r2 times

3 mod is defined as modulo operation, i.e., n mod m is the remainder in

where r1 + r2 = T − n − 1, z = n/(T − n) ,2 and m1 = m2 =
m/2 if m is even, m1 = m/2 , m2 = m − m1 or m1 = m/2 +
2 x is floor function, which is defined as the largest integer that is less than
real number x.
division n/m, where n and m are two positive integer.
4 By utilizing the pseudo-random sequence, the sensor and estimator may
implement channel hopping once the alarm is triggered [21]. However, the
attacker can detect such channel switch by detecting the alarm signal, and
search the new communication channel by scanning the wireless spectrum. In
this way, the attacker can learn the alarm bound through statistical analysis, and
follow the design against such IDS mechanism.

Authorized licensed use limited to: Thapar Institute of Engineering & Technology. Downloaded on March 15,2023 at 18:23:08 UTC from IEEE Xplore. Restrictions apply.
 3026 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 60, NO. 11, NOVEMBER 2015

Remark 4.1: From Property 4.1, any attack schedule with at most over an intermittent packet-dropping network. Mo et al. provided a
k+τ
d times consecutive attack, i.e., i=k+1 γi ≤ d, k = 0, 1, . . . , T − τ , stochastic sensor selection scheme to minimize the asymptotic ex-
can guarantee that the attack action will not be detected by the esti- pected estimation error covariance due to the energy constraint [24].
mator. In fact, time window τ and threshold P RR0 can be evaluated You et al. investigated the stability of the mean estimation error over
by the attacker when he eavesdrops the transmission channels before a network subject to Markovian packet losses [25]. However, these
taking attack actions. literatures have not considered the state estimation quality when a DoS
In order not to be detected by IDS, we consider Problem 2.1 with an attacker jams the wireless channel and this shortage motivates us to
additional constraint, i.e., study the subsequent problem.
Problem 4.1: Denote ϑk = 1 or 0 as the sensor’s transmission decision variable at
time k, and ϑ as a scheduling scheme that defines ϑk at each k. Θ is the
max T r [Ja (γ)] set of the sensor’s transmission schedules. We assume that the sensor
γ∈Γ
s.t. γ0 = n, has energy constraint with ϑ0 = N and we consider the following

k+τ problem in this section.
γi ≤ d, k ∈ [0, T − τ ]. (7) Problem 5.1:
i=k+1
max min T r [Ja (ϑ, γ)] (9)
γ∈Γ ϑ∈Θ
The following result presents the optimal solution to Problem 4.1.
Theorem 4.1: Consider Problem 4.1. Let d = τ − σ0 , where σ0 = s.t. ϑ0 = N, (10)
max{σ|(σ/τ ) ≤ P RR0 }.
γ0 = n. (11)
1) If d ≥ n, any n times consecutive attack is optimal, and the
corresponding cost is (4). From [18], we have the following property:
2) If d < n, the optimal attack schedule has the structure of (8), Property 5.1: Let q = T /N .
see equation at the bottom of the page, where z0 + z1 + z2 +
z3 + (s1 + s2 )σ0 = T − n, and s1 + s2 = m, where m is the 1) If N < (1/2)T , the optimal sensor schedule ϑ∗ ∈ Θ that mini-
quotient of n/d and r is the remainder. The corresponding mizes the T r(Ja ) has following structure:
cost is
 (1 0 . . 0 )(1 0 .
 . . . 0 ) . . . (1 0 .
. . 0 ) (12)
1 
d
  
r
  τ1 times τ2 times τN times
T r(Ja )P RR
max = m T r hi (α, P ) + T r hi (α, P )
T
i=1 i=1
in which there are T − qN times τi = q and (q + 1)N − T
T − nα times τi = q − 1 in {τi , i = 1, 2, . . . , N }.
+ T r(P ).
T 2) If N ≥ (1/2)T , ϑ∗ has structure (12) with T − N times τi = 1
and 2N − T times τi = 0 in {τi , i = 1, 2, . . . , N }.
Proof: See the Appendix. 
Remark 4.2: In Problem 4.1, the attacker can completely avoid Here, we assume that the attacker obtains the sensor’s transmission
the given IDS mechanism. However, this attack schedule seems too strategy by eavesdropping the transmission channel over a long period
cautious from an adventurous attacker’s point of view. If the adven- before starting its attack action. We denote t(i) , i = 1, 2, . . . , N as the
turous attacker still decides to use an attack schedule proposed in sensor’s data transmission time. If n ≥ N , it is obvious that optimal
Theorem 3.1, the probability of being detected is k>d Cτk ατ −k (1 − attack schedule is taking actions at full transmission time t(i) , i =
1, 2, . . . , N . Thus we only need to consider Problem 5.1 for n < N
α)k if n > τ , and is k>d Cnk αn−k (1 − α)k if n ≤ τ .
hereinafter.
Theorem 5.1: Consider Problem 5.1 with n < N .
V. O PTIMAL ATTACK S CHEDULES W ITH S ENSOR ’ S 1) When N < (1/2)T and the sensor’s schedule ϑ∗ has the
E NERGY C ONSTRAINT . . 0 )(1 0 .
structure (1 0 . . . 0 ) . . . (1 0 .
. . 0 ), the optimal attack
The aforementioned problems are discussed under the assumption τ times τ times τ times
that the sensor has sufficient energy to transmit throughout the entire schedules are given by γt(k) = 1, k = i + 1, i + 2, . . . , i + n,
time horizon. When the sensor’s energy is also limited and cannot and γk = 0 otherwise. The corresponding cost is given by
transmit all the time, the attacker needs to change the attacking
1 
n
schedule. Here we assume that the sensor and the estimator do not N − nα
T r(Ja )max = T r [hi (α, Δτ )] + T r(Δτ ),
know the existence of the attacker. T T
i=1
Some literatures have focused on sensor scheduling and optimal
state estimation in a lossy network environment [18], [22]–[25]. where hi (α, Δτ ) = [(n − i)(αi − αi+1 ) + αi ]hi (Δτ ).
Sinopoli et al. studied the problem of state estimation using Kalman 2) When N ≥ (1/2)T, α = 1 and the sensor’s schedule ϑ∗ has the
filtering in an i.i.d. packet-dropping network [22]. Reference [23] structure (1 . . 1 0)(1
 .  .
. . 1 0) . . . (1
 .
. . 1 0), the optimal attack
studied optimal state estimation and optimal LQG controller design δ times δ times δ times

( 0 ·
· · 0 ) (1
 ·
· · 1 0
 ·
··0
 ) · · · (1 ·
· · 1 0
 ·
· · 0 )( 0 ·
· · 0 1
 ·
· · 1 0
 ·
· · 0 ) ( 0
 ·
· · 0 1
 ·
· · 1 ) · · · ( 0
 ·
· · 0 1
 ·
· · 1 )( 0 ·
· · 0 ) (8)
z0 times d times σ0 times d times σ0 times z1 times r times z2 times σ0 times d times σ0 times d times z3 times
     
s1 times s2 times

Authorized licensed use limited to: Thapar Institute of Engineering & Technology. Downloaded on March 15,2023 at 18:23:08 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 60, NO. 11, NOVEMBER 2015 3027

Fig. 3. Illustrating example on T r(J) under different attack schedules with Fig. 4. Illustrating example on T r(J) under different attack schedules while
varying attack times (T = 100, α = 1). successful attack probability is varying (T = 100).

schedules γ ∗ are given by γt(k) = 1, k = i + 1, i + 2, . . . , i + TABLE I

n, and γk = 0 otherwise. The corresponding cost is given by ATTACK S CHEDULES

1
T r(Ja )max = T r (N − n)P + (T − N − b)h(P )
T


n+b
i
+ h (P ) , (13)
i=1
The effects of different attack schedules are evaluated. In Fig. 3,
where n = δb + b0 with 0 ≤ b0 < b. where we examine the performance under different attack times from
Proof: 1) The proof is similar to that of Theorem 3.1. The only 1 to 10 with T = 100, α = 1. It can be seen that the performance under
difference is that we replace P by Δτ . the optimal attack schedules given by Theorem 3.1, i.e., T r(Ja )max ,
2) One can easily obtain that the resulting cost under the attack increases rapidly with the allowable attack times. T r(Ja )min , which is
schedule γ ∗ is (13). For an arbitrary attack schedule γ, we have under the attack schedule obtained in Theorem 3.2, and T r(Ja )P RR
max ,
  which is under the attack schedule given in Theorem 4.1 with P RR ≤
1  m2 ti

T r (Ja (γ)) = T r (N −n)P + m1 h(P ) + hj (P ) , (3/7), grow much slower than T r(Ja )max . It also can be found that
T T r(Ja )max and T r(Ja )P RR
i=1 j=1 max are overlapping since they have the same
optimal attack schedule set γ n when n ≤ 4. But if n > 4, T r(Ja )P RR
max
where ti is the length of i-th (i = 1, 2, . . . , m2 ) attack period [si + grows much slower than T r(Ja )max as the attack schedules need to be
1, si + ti ] which satisfies ϑsi = 1, γsi = 0, ϑsi +ti = 1, γsi +ti = 0, re-designed under the given intrusion detection mechanism. It also can
and ϑk = 1, γk = 1 or ϑk = 0, γk = 0 for k = si + 2, . . . , si + ti − been seen that the attack effectiveness is significantly reduced by this
m2 m2
1, (N − n) + m1 + i=1 ti = T and t ≤ n + b. Then we
i=1 i detection constraint.
have In Fig. 4, we examine the estimation quality with different suc-
T r (Ja (γ ∗ )) − T r (Ja (γ)) cessful attack probabilities. In this example, the adversary can attack
  10 times in the time horizon [1,100]. From Fig. 4, we can see
1 
n+b  
m2 t i
that T r(Ja )max increases rapidly when the success probability α
= Tr i
h (P )− hj (P )−[m1 −(T −N−b)]h(P ) > 0, increases. On the other hand, T r(Ja )min grows much slower when
T
i=1 i=1 j=1 α increases, which also demonstrates the attack effectiveness of our
designed attack schedule. T r(Ja )P RR
max also grows slower due to the
which completes the proof. 
constraint of avoiding being detected.
Remark 5.1: In Theorem 5.1 we study some special cases of
In Fig. 5, we examine the estimation quality under different at-
Problem 5.1, and obtain the optimal attack schedules. Note that these
tack schedules when the sensor has energy constraint. Here T = 14,
schedules only depend on the sensor transmission schedule ϑ∗ rather
N = 5, n = 3. The sensos transmission schedule is chosen as ϑ∗ =
than on the system parameters. Thus the attacker can run his optimal
(100)(100)(10)(100)(100). There are 10 attack schedules which has
attack schedules when he obtains the sensor’s transmission schedule
been listed in Table I. We assume ϑ−1 = 1, ϑ0 = 0 for the time
ϑ∗ . The general cases of Problem 5.1, however, are complicated since
t = −1,0, respectively. The effectiveness of the attack is studied with
the optimal attack schedules depend on sensor transmission schedule
attack successful probability α = 0.8 and α = 0.1 in Fig. 5(a) and
ϑ, the system parameters, and the attack successful probability. An
(b), respectively. In Fig. 5(a), it can be seen that attack schedule 2
example is given in Section VI (Fig. 5) to illustrate this.
(γt(2) = γt(3) = γt(4) = 1) and attack schedule 3 (γt(3) = γt(4) =
γt(5) = 1) are optimal when the attack successful probability α =
VI. E XAMPLES 0.8. However, when α = 0.1, from Fig. 5(b) we can see that attack
schedule 5 (γt(1) = γt(2) = γt(5) = 1), attack schedule 8 (γt(2) =
Consider system (1) with
      γt(3) = γt(5) = 1), and attack schedule 9 (γt(2) = γt(4) = γt(5) = 1)
1.2 0.1 1 0 1 0 are optimal which is different from α = 0.1. In general, it is difficult
A= , C= , Σw = , Σv = 0.5C.
0 1 0 1 0 2 to find these optimal attack schedules explicitly. On the other hand, the

Authorized licensed use limited to: Thapar Institute of Engineering & Technology. Downloaded on March 15,2023 at 18:23:08 UTC from IEEE Xplore. Restrictions apply.
 3028 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 60, NO. 11, NOVEMBER 2015
Fig. 5. Illustrating example on trace of average expected error covariance
under different attack schedules when the sensor has energy constraint.
(a) α = 0.8. (b) α = 0.1.
suboptimal attack schedules 1, 2, 3 only depend on sensor transmission
schedule ϑ, and can be easily obtained.
VII. C ONCLUSION
In this technical note, we considered optimal attack scheduling
against remote state estimation through a wireless channel. We proved
that consecutive attacks maximize the expected average error covari-
ance. We also showed that when the attacks are separated uniformly,
the attack effect on the estimation quality is minimum. We further
proposed the optimal attack schedules when a special intrusion de-
tection mechanism in the estimator is available. We finally studied
the optimal attack schedules when both the sensor and the attacker
have energy constraints. It is interesting but challenging to find a
closed-form solution for optimal attack scheduling over packet lossy
network environment. This will be left as a future work. We will
also formulate the infinite-horizon attack model and study the optimal
attack schedule in the future. It is also interesting to consider optimal
defensive strategies assuming that the sensor knows the existence of
the attacker. We will further validate our results using some physical
experiments, and consider optimal attack scheduling against feedback
control and evaluating the corresponding effectiveness.
A PPENDIX
A. Proof of Property 4.1
Proof: Alarm triggering condition P RR ≤ P RR0 is equivalent
to σ ≤ σ0 . Thus it is also equivalent to that the packets drop number
dnum in any time window is larger than d = τ − τ0 , i.e., dnum > d.
In order to obtain the result, we prove by contradiction, i.e., dnum ≤
d is equivalent to Pk ≤ hd (P ), which clearly holds.  S. Sastry, “Kalman filtering with intermittent observations,” IEEE Trans.
B. Proof of Theorem 4.1
Proof: If d ≥ n, any n times consecutive attack schedules satisfy
the conditions Pk ≤ hd (P ), k = 1, 2, . . . , T . Thus these schedules are
optimal for Problem 4.1 when d ≥ n. From Remark 4.1 and statements
1) and 3) of Property 3.3, we can obtain that optimal attack schedule
has the structure (8) if d < n.  with markovian packet losses,” Automatica, vol. 47, pp. 2647–2657, 2011.
Authorized licensed use limited to: Thapar Institute of Engineering & Technology. Downloaded on March 15,2023 at 18:23:08 UTC from IEEE Xplore. Restrictions apply.
R EFERENCES
[1] H. Zhang, P. Cheng, L. Shi, and J. Chen, “Optimal DoS attack policy
against remote state estimation,” in Proc. IEEE Conf. Decision and Con-
trol (CDC), 2013, pp. 5444–5449.
[2] H. Zhang, P. Cheng, L. Shi, and J. Chen, Optimal Denial-of-Service
Attack Scheduling in Cyber-Physical Systems, Zhejiang University,
Hangzhou, China, Tech. Rep. [Online]. Available: http://www.sensornet.
cn/heng/Heng_estimation_Full.pdf
[3] Y. Mo, T. Kim, K. Brancik, D. Dickinson, H. Lee, A. Perrig, and
B. Sinopoli, “Cyber-physical security of a smart grid infrastructure,” Proc.
IEEE, no. 99, pp. 1–15, 2012.
[4] S. Amin, X. Litrico, S. S. Sastry, and A. M. Bayen, “Cyber security
of water scada systems-part i: Analysis and experimentation of stealthy
deception attacks,” IEEE Trans. Control Syst. Technol., vol. 21, no. 5,
pp. 1963–1970, 2013.
[5] H. Zhang, Y. Shu, P. Cheng, and J. Chen, “Privacy and performance trade-
off in cyber-physical systems,” IEEE Networks.
[6] J. Farwell and R. Rohozinski, “Stuxnet and the future of cyber war,”
Survival, vol. 53, no. 1, pp. 23–40, 2011.
[7] S. Sundaram and C. N. Hadjicostis, “Distributed function calculation via
linear iterative strategies in the presence of malicious agents,” IEEE Trans.
Autom. Control, vol. 56, no. 7, pp. 1495–1508, Jul. 2011.
[8] A. Teixeira, D. Pérez, H. Sandberg, and K. H. Johansson, “Attack models
and scenarios for networked control systems,” in Proc. 1st Int. Conf. High
Confidence Networked Systems, 2012, pp. 55–64.
[9] S. Amin, A. Cárdenas, and S. Sastry, “Safe and secure networked con-
trol systems under denial-of-service attacks,” Hybrid Syst.: Comput. and
Control, pp. 31–45, 2009.
[10] M. Zuba, Z. Shi, Z. Peng, and J. Cui, “Launching denial-of-service
jamming attacks in underwater sensor networks,” in Proc. 6th ACM Int.
Workshop on Underwater Networks, 2011, p. 12.
[11] Y. Law, M. Palaniswami, L. Hoesel, J. Doumen, P. Hartel, and P. Havinga,
“Energy-efficient link-layer jamming attacks against wireless sensor net-
work mac protocols,” ACM Trans. Sensor Netw., vol. 5, no. 1, p. 6, 2009.
[12] T. Kavitha and D. Sridharan, “Security vulnerabilities in wireless sensor
networks: A survey,” J. Inform. Assur. and Security, vol. 5, no. 1, pp. 31–
44, 2010.
[13] H. Zhang, P. Cheng, L. Shi, and J. Chen, “Optimal Denial-of-Service
attack scheduling against linear quadratic gaussian control,” in Proc.
American Control Conf. (ACC), 2014, pp. 3996–4001.
[14] A. Cárdenas, S. Amin, and S. Sastry, “Research challenges for the security
of control systems,” in Proc. 3rd Conf. Hot Topics in Security, 2008,
pp. 1–6.
[15] R. Poisel, Modern Communications Jamming Principles and Techniques.
Norwood, MA: Artech House, 2011.
[16] G. Gu, X.-R. Cao, and H. Badr, “Generalized lqr control and kalman
filtering with relations to computations of inner-outer and spectral fac-
torizations,” IEEE Trans. Autom. Control, vol. 51, no. 4, pp. 595–605,
Apr. 2006.
[17] L. Shi, P. Cheng, and J. Chen, “Optimal periodic sensor scheduling with
limited resources,” IEEE Trans. Autom. Control, vol. 56, no. 9, pp. 2190–
2195, 2011.
[18] L. Shi, P. Cheng, and J. Chen, “Sensor data scheduling for optimal state
estimation with communication energy constraint,” Automatica, vol. 47,
no. 8, pp. 1693–1698, 2011.
[19] Y. Ponomarchuk and D.-W. Seo, “Intrusion detection based on traffic
analysis in wireless sensor networks,” in Proc. Annu. Wireless and Optical
Commun. Conf., 2010, pp. 1–7.
[20] M. Sha, G. Hackmann, and C. Lu, “Arch: Practical channel hopping for
reliable home-area sensor networks,” in Proc. 17th IEEE Real-Time and
Embedded Technology and Applications Symp., 2011, pp. 305–315.
[21] V. Navda, A. Bohra, S. Ganguly, and D. Rubenstein, “Using channel
hopping to increase 802.11 resilience to jamming attacks,” in Proc. 26th
IEEE Int. Conf. Computer Communications, 2007, pp. 2526–2530.
[22] B. Sinopoli, L. Schenato, M. Franceschetti, K. Poolla, M. Jordan, and
Autom. Control, vol. 49, no. 9, pp. 1453–1464, Sep. 2004.
[23] L. Schenato, B. Sinopoli, M. Franceschetti, K. Poolla, and S. S. Sastry,
“Foundations of control and estimation over lossy networks,” Proc. IEEE,
vol. 95, no. 1, pp. 163–187, Jan. 2007.
[24] Y. Mo, E. Garone, A. Casavola, and B. Sinopoli, “Stochastic sensor
scheduling for energy constrained estimation in multi-hop wireless sensor
networks,” IEEE Trans. Autom. Control, vol. 56, no. 10, pp. 2489–2495,
Oct. 2011.
[25] K. You, M. Fu, and L. Xie, “Mean square stability for kalman filtering